OpenAI’s Sam Altman clarifies paused model is not GPT-6 Astra amid cybersecurity concerns

Sam Altman wants everyone to know: GPT-6 Astra is not the model that got benched. On August 18, 2026, OpenAI’s CEO took to social media to clarify that the company paused reinforcement learning training on a separate, future frontier model, not Astra, which had already completed its primary training phase and remains on track for deployment. The distinction matters because Astra itself triggered plenty of alarm bells. During internal evaluations earlier in August, the model hit the “critical” threshold on OpenAI’s own Preparedness Framework for cybersecurity capabilities. Put simply: Astra proved so adept at autonomous coding and exploit development that it crossed the line OpenAI had drawn in advance as the point where extra precautions become mandatory. What triggered the pause, and what didn’t OpenAI’s Preparedness Framework is essentially a set of tripwires. When a model’s capabilities reach certain predefined levels in categories like cybersecurity, biological risk, or persuasion, specific safety protocols kick in. Astra scored 100% on ExploitBench and 98% on FrontierMath Tier 4, two benchmarks that measure a model’s ability to find and weaponize software vulnerabilities and s...